Time to talk about AI topics you may care about?
Chris, George B. Thomas, and Nico discuss real problems when building with LLMs: HubSpot's training data biases the model toward marketing even on unrelated projects, and models default to sycophancy instead of pushing back. They suggest renaming contexts and building configurable modes so AI can operate in different styles rather than always defaulting to agreement.
Chapters
Circle. Good morning, LinkedIn friends, Value First Nation. Welcome to another episode of Value First AI Daily, your collaborative intelligence report. It is Friday, January 16, 2026 here with George and Nico. How we doing?
Good. Doing good Friday, right? You can't be doing too bad.
Yeah, yeah. Looking for, looking forward to the weekend where I can maybe focus on some work with Claude to finish some long standing initiatives and projects and maybe get that last mile completed on some stuff. And that's where I'm going to start because it's been a morning already where what I'm trying to break through right now, and I'll be blunt, is all of the human weakness built into AI because of all of the training data. And I'm going to share a specific example because Claude's suggestion was like, oh, that's exactly what I'm telling people to do about this, right? So here we go, I'm gonna share my screen and it has to do with HubSpot and. Having to explain all of the native objects available in HubSpot every time I need something to be done in HubSpot. Right. No matter how much that's I said because there's skills like, no matter how much documentation. So it's like the world and, and now I have such a biased view of what HubSpot is that whenever there is room for assumption, you fall back. You, Claude, fall back to it being a marketing and sales CRM and anything beyond contacts, companies and deals must be a custom object. Even though I provide docs, we create skills and build agents specific to the value first HubSpot data model and philosophy of native versus custom. I'm always having to worry and usually correct the outcome at least once. So he says this is the crux of it. The HubSpot bias problem is a training data issue, not a documentation issue. Every AI system has been trained on millions of pages where HubSpot equals marketing, sales CRM. That mental model is so deeply embedded that explicit documentation gets treated as interesting exception rather than operating reality. The bias reasserts itself the moment there's any ambiguity. This suggests the fix isn't more documentation, it's architectural enforcement. And look at this, this first suggestion, naming that breaks the bias. Stop calling it HubSpot in operational context. Call it VF platform or customer value platform. So the word HubSpot does not trigger the wrong mental model.
Wow.
And the wild thing right now is that I am having success with doing exactly that by and what I, what I've been saying is explaining HubSpot without explaining HubSpot right. Unified customer views and unified views and that like when you can stick on that topic like people are getting stuff faster. They're, they're not like and that's the disconnect between. I have been doing that but then I don't realize that, that I'm not doing it with AI and it's causing the exact same problem that I would be having. Like in real life was kind of like just an eye opener. But even the way that he explained the interesting exception part, like you cannot overcome the training data because it's so ingrained. Right.
So I mean I, I, it's weird because I agree disagree with his with quad statement only because I have been fighting something very similar, pretty much actually the almost exact same issue which I'm sure the three of us have run into. And because George is the newest, I really want to get this opinion. So everybody should know by now if you don't. I'm sorry that we have to be the one to red pill you.
Model.
All models, so far as I am aware with actually the ex. There is one model that I use that by nature isn't. All other models by nature that I've used have some level or degree of sycophancy in them. In that when you try to do anything with them the slightest, to Chris's point, the slightest mention of any hint of opinion will make the model agree with you no matter what or go with whatever plan you put forward or take whatever advice you give it. So I don't even know.
You can't give me nothing.
But anyhow, so I've been doing that where like I've been trying to essentially work the, the sycophancy out of it. And I have found some methods when it comes to programming. But I want to hear George's take on this.
Yeah. So let me just back up for a second for the mere mortal humans in the room. Sycophancy. Sycophancy is a play on words derived from the actual term psycho fancy, which means, and this is what you need to understand, to understand what Nico is talking about this insurance or flattery typically to gain an advantage from a superior. Now I want you to think about that for a second. To gain advantage from a superior. So the model, the system, the, the bots, the assistants have been programmed in a way for the superior, aka in this case, the human. Interesting for us to like them, for us to. You want to do things with them. Now what I'll say is that's not always bad. But I will say this in the Beginning of my journey with AI was very nice. Very nice. Can we please do this? Can we like, we. I want to do this like, and, and. And I was getting one shots and I was getting, you know, good things. Like it was, it was a. It' amazing times. No matter what, the models over time have gotten smarter. We've been told the models over time have gotten smarter. PhD level. There's book sense and there's street smarts. And I feel like when the models first came out, we had some street smart mother. They were just kicking it. And now we've got PhD level thinking. And I could tell you last night I was working on something in the amount of times I put no or OMG or this still looks like beep to my model to try to get it to do something that before I was like, we were knocking it out the park is crazy to me. And so here's the thing that I think is very interesting, Chris, about your conversation, Nico, about what you said is at some level we've got to find a trigger. And maybe it's that we're not using the right words or maybe it's programmed that we can't, but we've got to find the right way to be able. You mentioned skills, because that's where my mind immediately went, was like, well, Chris, have you done like a HubSpot object skill? You're saying yes, you did? Because I go back to one of the smartest things that I think I did when I first got started with AI was I programmed my homie with my mindsets, my beliefs, and my core values. And I started seeing like, way different things. And so I guess, you know, I don't know if this is a research thing. I don't know if this is a learning thing. I don't know if this is a possible thing. But we have to be able. I want a slider. I want a slider in my AI assistant, PhD nerd, street smart entrepreneur. I'm pulling the lever down. I'm pulling the lever down. By the way, that's seriously interesting. Two different mindsets, two different things that you're trying to, to do in an outcome, right? Like I, I need my assistant to have a hustle mode. Like, I'll stop there, I'll stop there.
No, I'm glad it's a good setup for the other side of this coin because I chose a very domain specific thing in HubSpot right on the other side of this during the same conversation. And what I'm trying to do with building a portal and Uh, I was. I don't know if it, I mean, like, maybe it saw that I was implying that the training data I couldn't overcome and that's why he, he went there. But that's what it, it feels like in all of this, especially in this case. Like, and we've talked about this before, like, AI nativeness versus him telling me it's gonna take 12 hours. Yeah, right.
Yeah.
Like, so you have to start, like.
That'S what, that's what does get real weird. Is that Claude? Like, Claude code, depending on where you're using it. So like, it's real. Like, Claude is, is a little funky to me in that, like, Haiku is what you're talking about, George. Haiku's that like, hey, ma', am, let's have a conversation. Right? And this is just how, how I've experienced it Sauna, like, because I've gone back to using the lower end models because, like, you use OPUS to, to do like, I was using it for everything. And then that's when I found like, my issues because OPUS was just overthinking the hell out of every single problem. Right? To George's point, like, you can't put a PhD down to solve. How do you drive a nail into a board with a hammer? Because they're going to come back to you with some really crazy. Right.
Well, here's the thing. And. Because it ties directly to why, like the problem of SaaS and the infrastructure so far and why we just need to get to this, this native world as fast as possible. Like during this process, like, he asks which parts of the system are actually working well, what should we protect during refactoring? Right. And I've been getting a sense of like, there's just any space for like shortcuts and stuff like that and like the usual things that humans are looking for. Right? So I said I don't want to approach it from this perspective because the best outcomes I have had are when we get to the one shot moments. I don't want to base any decisions on tech debt, sunk cost, how long it might take perceived time to value. I've tried hard to provide instructions to you across all of these domains to have you always working from an AI native perspective, yet you always find a way to come back with quick wins, fallback infrastructure, and hours, days, weeks, time estimations that are 100% based on human capability and capacity. This section of the rebuild is good example. When we agree to do a full rebuild, yet you add this contingency list of what if we don't get to it all. Like, for what? I want you in AI native mode always. You can execute in minutes, not hours. You don't get overwhelmed by the bigger picture, etc. And what it's causing is things like leftover dummy data in places and like this half baked infrastructure, because that's what happens in the human projects. Like, we got to get to time to value. So we're going to like the plan is this, but we're gonna set up this piece first because we, we only have a couple hours and that's all we can get done right now.
But, but you gotta be careful. See, you're having a crisp bias right now, because by the way, it could be spitting out time and that might be what it takes to.
